Homepage
The formatted source code for this file is here.
And a raw version here.
Previous work by Youngser Park can be found here.
Following from previous pages, this page will focus on filtering the data before clustering to explore if filtering improves the outcome of clustering.
Here we read in the data and select a random half of it for exploration.
featFull <- fread("../data/synapsinR_7thA.tif.Pivots.txt.2011Features.txt",showProgress=FALSE)
locFull <- fread("../data/synapsinR_7thA.tif.Pivots.txt",showProgress=FALSE)
### Setting a seed and creating an index vector
### to select half of the data
set.seed(2^10)
half1 <- sample(dim(featFull)[1],dim(featFull)[1]/2)
half2 <- setdiff(1:dim(featFull)[1],half1)
feat <- featFull[half1,]
loc <- locFull[half1,]
dim(feat)# [1] 559649 144
## Setting the channel names
channel <- c('Synap_1','Synap_2','VGlut1_t1','VGlut1_t2','VGlut2','Vglut3',
'psd','glur2','nmdar1','nr2b','gad','VGAT',
'PV','Gephyr','GABAR1','GABABR','CR1','5HT1A',
'NOS','TH','VACht','Synapo','tubuli','DAPI')
## Setting the channel types
channel.type <- c('ex.pre','ex.pre','ex.pre','ex.pre','ex.pre','in.pre.small',
'ex.post','ex.post','ex.post','ex.post','in.pre','in.pre',
'in.pre','in.post','in.post','in.post','in.pre.small','other',
'ex.post','other','other','ex.post','none','none')
nchannel <- length(channel)
nfeat <- ncol(feat) / nchannel
## Createing factor variables for channel and channel type sorted properly
ffchannel <- (factor(channel.type,
levels= c("ex.pre","ex.post","in.pre","in.post","in.pre.small","other","none")
))
fchannel <- as.numeric(factor(channel.type,
levels= c("ex.pre","ex.post","in.pre","in.post","in.pre.small","other","none")
))
ford <- order(fchannel)
## Setting up colors for channel types
Syncol <- c("#197300","#5ed155","#660000","#cc0000","#ff9933","mediumblue","gold")
ccol <- Syncol[fchannel]
exType <- factor(c(rep("ex",11),rep("in",6),rep("other",7)),ordered=TRUE)
exCol<-exType;levels(exCol) <- c("#197300","#990000","mediumblue");
exCol <- as.character(exCol)
fname <- as.vector(sapply(channel,function(x) paste0(x,paste0("F",0:5))))
names(feat) <- fname
fcol <- rep(ccol, each=6)
mycol <- colorpanel(100, "purple", "black", "green")
mycol2 <- matlab.like(nchannel)f <- lapply(1:6,function(x){seq(x,ncol(feat),by=nfeat)})
featF <- lapply(f,function(x){subset(feat,select=x)})
featF0 <- featF[[1]]
f01e3 <- 1e3*data.table(apply(X=featF0, 2, function(x){((x-min(x))/(max(x)-min(x)))}))
fs <- f01e3
### Taking log_10 on data with 0's removed
ans <- apply(featF0, 1, function(row){ any(row == 0)})
logF0 <- log10(featF0[!ans,])
slogF0 <- logF0[,lapply(.SD,scale, center=TRUE,scale=TRUE)]We now have the following data sets:
featF0: The feature vector looking only at the integrated brightness features.fs: The feature vector scaled between \([0,1000]\).logF0: The feature vector, with 0’s removed, then \(log_{10}\) is applied.slogF0: The feature vector, with 0’s removed, then \(log_{10}\), then scaled by subtracting the mean and dividing by the sample standard deviation.df1 <- melt(as.matrix(fs))
names(df1) <- c("ind","channel","value")
df1$type <- factor(rep(ffchannel,each=dim(fs)[1]),levels=levels(ffchannel))
lvo <- c(1:5,7:10,19,22,11:16,6,17,18,20,21,23,24)
levels(df1$channel)<-levels(df1$channel)[lvo]
ts <- 22
gg1 <- ggplot(df1, aes(x=value)) +
scale_color_manual(values=ccol[lvo]) +
scale_fill_manual(values=ccol[lvo]) +
geom_histogram(aes(y=..density..,group=channel,colour=channel),bins=100) +
geom_density(aes(group=channel, color=channel),size=1.5) +
facet_wrap( ~ channel, scale='free', ncol=6) +
theme(plot.title=element_text(size=ts),
axis.title.x=element_text(size=ts),
axis.title.y=element_text(size=ts),
legend.title=element_text(size=ts),
legend.text=element_text(size=ts-2),
axis.text=element_text(size=ts-2),
strip.text=element_text(size=ts),
legend.position='none')+
ggtitle("Kernel Density Estimates of `fs` data.")
print(gg1)fs data.cmatfs <- cor(fs)
corrplot(cmatfs,method="color",tl.col=ccol[ford], tl.cex=1)pcaf0 <- prcomp(featF0,scale=TRUE, center=TRUE)
pcafs <- prcomp(fs,scale=FALSE, center=FALSE)
elpcaf0 <- getElbows(pcaf0$sdev, plot=FALSE)
elpcafs <- getElbows(pcafs$sdev, plot=FALSE)We run a Hierachical K-means++ for \(K=2\) on the fs data with 4 levels.
set.seed(2^13)
L <- bhkmpp(fs,blevels=4)corkp1 <- cor(fs[L[[1]] == 1,])
corkp2 <- cor(fs[L[[1]] == 2,])
par(mfrow=c(2,2))
corrplot(corkp1,method="color",tl.col=ccol[ford], tl.cex=0.8)
corrplot(corkp2,method="color",tl.col=ccol[ford], tl.cex=0.8)
corrplot(sqrt((corkp1 - corkp2)^2),method="color",tl.col=ccol[ford], tl.cex=0.8)Notice that the non-synaptic markers change very little between clusters. Also note that the correlations between (gad, VGAT, PV, Gephyr) and VGlut1 at both times change significantly between clusters.
## Formatting data for heatmap
aggp <- aggregate(fs,by=list(lab=L[[1]]),FUN=mean)
aggp <- as.matrix(aggp[,-1])
rownames(aggp) <- clusterFraction(L[[1]])The following are heatmaps generated from clustering via K-means++ (at level 1)
heatmap.2(as.matrix(aggp),dendrogram='row',Colv=NA,trace="none", col=mycol,colCol=ccol[ford],cexRow=0.8, keysize=1.25,symkey=FALSE,symbreaks=FALSE,scale="none", srtCol=90,main="Heatmap of `fs` data.") # [1] "#197300" "#197300" "#197300" "#197300" "#197300"
# [6] "#5ed155" "#5ed155" "#5ed155" "#5ed155" "#5ed155"
# [11] "#5ed155" "#660000" "#660000" "#660000" "#cc0000"
# [16] "#cc0000" "#cc0000" "#ff9933" "#ff9933" "mediumblue"
# [21] "mediumblue" "mediumblue" "gold" "gold"
Percentage of data within cluster is presented on the right side of the heatmap.
Here we look at the kernel density estimates within each cluster to compare.
df2 <- melt(as.matrix(fs))
names(df2) <- c("ind","channel","value")
df2$cluster <- L[[1]]
df2$type <- factor(rep(ffchannel,each=dim(fs)[1]),levels=levels(ffchannel))
gg2 <- ggplot(df2, aes(x=value)) +
scale_colour_manual(values=ccol) +
scale_x_continuous(limits=c(0,400)) +
geom_histogram(aes(y=..density..,group=channel,colour=channel),bins=250) +
geom_density(aes(group=channel, color=channel),size=1.5) +
facet_grid(channel ~ cluster, scale='free') +
theme(strip.text.y=element_text(angle=0)) +
#guides(col=guide_legend(ncol=1))
theme(strip.text.y=element_text(angle=0),
plot.title=element_text(size=ts),
axis.title.x=element_text(size=ts),
axis.title.y=element_text(size=ts),
legend.title=element_text(size=ts),
legend.text=element_text(size=ts-2),
strip.text=element_text(size=ts),
legend.position='none')print(gg2)fs data given cluster from km++ level 1Using the location data and the results of K-means++ we show a 3d scatter plot colored accoding to cluster.
set.seed(2^12)
s1 <- sample(dim(loc)[1],5e4)
locs1 <- loc[s1,]
locs1$cluster <- L[[1]][s1]
plot3d(locs1$V1,locs1$V2,locs1$V3,
col=brewer.pal(4,"Set1")[order(table(L[[1]]))][locs1$cluster],
alpha=0.75,
xlab='x',
ylab='y',
zlab='z')
subid <- currentSubscene3d()
rglwidget(elementId="plot3dLocations", height=720, width=720)corLV2 <- lapply(c(1:4),function(x){cor(fs[L[[2]] == x,])})
par(mfrow=c(2,2))
for(pl in corLV2){
corrplot(pl,method="color",tl.col=ccol[ford], tl.cex=0.8)
}## Formatting data for heatmap
aggp2 <- aggregate(fs,by=list(lab=L[[2]]),FUN=function(x){mean(x)})
aggp2 <- as.matrix(aggp2[,-1])
rownames(aggp2) <- clusterFraction(L[[2]])The following are heatmaps generated from clustering via K-means++
heatmap.2(as.matrix(aggp2),dendrogram='row',Colv=NA,trace="none", col=mycol,colCol=ccol[ford],cexRow=0.8, keysize=1.25,symkey=FALSE,symbreaks=FALSE,scale="none", srtCol=90,main="Heatmap of `fs` data.") # [1] "#197300" "#197300" "#197300" "#197300" "#197300"
# [6] "#5ed155" "#5ed155" "#5ed155" "#5ed155" "#5ed155"
# [11] "#5ed155" "#660000" "#660000" "#660000" "#cc0000"
# [16] "#cc0000" "#cc0000" "#ff9933" "#ff9933" "mediumblue"
# [21] "mediumblue" "mediumblue" "gold" "gold"
Percentage of data within cluster is presented on the right side of the heatmap.
Here we look at the kernel density estimates within each cluster to compare.
df2 <- melt(as.matrix(fs))
names(df2) <- c("ind","channel","value")
df2$cluster <- L[[2]]
df2$type <- factor(rep(ffchannel,each=dim(fs)[1]),levels=levels(ffchannel))
gg3 <- ggplot(df2, aes(x=value)) +
scale_colour_manual(values=ccol) +
#scale_x_continuous(limits=c(0,400)) +
geom_histogram(aes(y=..density..,group=channel,colour=channel),bins=250) +
geom_density(aes(group=channel),size=.5, color='black', alpha=0.5) +
facet_grid(channel ~ cluster, scale='free') +
#guides(col=guide_legend(ncol=1))
theme(strip.text.y=element_text(angle=0),
plot.title=element_text(size=ts),
axis.title.x=element_text(size=ts),
axis.title.y=element_text(size=ts),
legend.title=element_text(size=ts),
legend.text=element_text(size=ts-2),
strip.text=element_text(size=ts),
legend.position='none')print(gg3)fs data given cluster from km++ level 2Using the location data and the results of K-means++ we show a 3d scatter plot colored according to cluster.
set.seed(2^12)
s1 <- sample(dim(loc)[1],5e4)
locs2 <- loc[s1,]
locs2$cluster <- L[[2]][s1]
YlOrBr <- c("#FFFFD4", "#FED98E", "#FE9929", "#D95F0E", "#993404")
col.pal <- colorRampPalette(YlOrBr)
plot3d(locs2$V1,locs2$V2,locs2$V3,
#col=colorpanel(8,"brown","blue")[order(table(L[[2]]))][locs2$cluster],
col=col.pal(8)[-seq(1,8,2)][order(table(L[[2]]))][locs2$cluster],
alpha=0.75,
xlab='x',
ylab='y',
zlab='z'
)
subid <- currentSubscene3d()
rglwidget(elementId="plot3dLocationsLV2")corLV3 <- lapply(c(1:8),function(x){cor(fs[L[[3]] == x,])})
par(mfrow=c(3,2))
for(pl in corLV3){
corrplot(pl,method="color",tl.col=ccol[ford], tl.cex=0.8)
}## Formatting data for heatmap
aggp3 <- aggregate(fs,by=list(lab=L[[3]]),FUN=function(x){mean(x)})
aggp3 <- as.matrix(aggp3[,-1])
rownames(aggp3) <- clusterFraction(L[[3]])The following are heatmaps generated from clustering via K-means++
heatmap.2(as.matrix(aggp3),dendrogram='row',Colv=NA,trace="none", col=mycol,colCol=ccol[ford],cexRow=0.8, keysize=1.25,symkey=FALSE,symbreaks=FALSE,scale="none", srtCol=90,main="Heatmap of `fs` data.") # [1] "#197300" "#197300" "#197300" "#197300" "#197300"
# [6] "#5ed155" "#5ed155" "#5ed155" "#5ed155" "#5ed155"
# [11] "#5ed155" "#660000" "#660000" "#660000" "#cc0000"
# [16] "#cc0000" "#cc0000" "#ff9933" "#ff9933" "mediumblue"
# [21] "mediumblue" "mediumblue" "gold" "gold"
Percentage of data within cluster is presented on the right side of the heatmap.
Here we look at the kernel density estimates within each cluster to compare.
df3 <- melt(as.matrix(fs))
names(df3) <- c("ind","channel","value")
df3$cluster <- L[[3]]
df3$type <- factor(rep(ffchannel,each=dim(fs)[1]),levels=levels(ffchannel))
gg4 <- ggplot(df3, aes(x=value)) +
scale_colour_manual(values=ccol) +
#scale_x_continuous(limits=c(0,400)) +
geom_histogram(aes(y=..density..,group=channel,colour=channel),bins=250) +
geom_density(aes(group=channel),size=.5, color='black', alpha=0.5) +
facet_grid(channel ~ cluster, scale='free') +
#guides(col=guide_legend(ncol=1)) +
theme(strip.text.y=element_text(angle=0),
plot.title=element_text(size=ts),
axis.title.x=element_text(size=ts),
axis.title.y=element_text(size=ts),
legend.title=element_text(size=ts),
legend.text=element_text(size=ts-2),
strip.text=element_text(size=ts),
legend.position='none')print(gg4)fs data given cluster from km++ level 3Using the location data and the results of K-means++ we show a 3d scatter plot colored according to cluster.
set.seed(2^12)
s1 <- sample(dim(loc)[1],5e4)
locs3 <- loc[s1,]
locs3$cluster <- L[[3]][s1]
plot3d(locs3$V1,locs3$V2,locs3$V3,
col=col.pal(16)[-seq(1,8,2)][order(table(L[[3]]))][locs3$cluster],
alpha=0.65,
xlab='x',
ylab='y',
zlab='z'
)
subid <- currentSubscene3d()
rglwidget(elementId="plot3dLocationsLV3")GABABR## re-formatting data for use in lattice
d1gab <- data.table(stack(fs, select=-GABABRF0))[,.(values)]
d1gab$GABABR <- fs$GABABRF0
### Adding relationship factor variables
nd <- paste0("GABABR","~",abbreviate(channel[-which(channel=="GABABR")]))
d1gab$ind <- factor(rep(nd,each=dim(fs)[1]),ordered=TRUE,levels=nd)
names(d1gab) <- c("x","y","g")gg5 <- ggplot(data=d1gab,aes(x=x,y=y, group=g)) +
geom_point(pch='.',alpha=0.2) +
geom_hex(bins=100) +
geom_smooth(method='lm',colour='red', alpha=0.7)+
facet_wrap( ~ g, scales='free_x') print(gg5)